How to Quantify AI Explainability for Improved Safety and Trust

Avi Rosenfeld

26 January 2023

3:00 pm - 4:30 pm

Please note that this event has now passed

In this talk we outline several ground-breaking directions for how explainable artificial intelligence (XAI) can be quantified and applied for improved AI safety and trust. We focus on an open challenge to properly match the need that motivates creating XAI and the algorithms used to provide that solution. To facilitate comparison between potential XAI solutions, we developed four metrics to quantify their algorithmic differences based on the number of the features in its input, rules it outputs, performance differences between algorithms, and stability within different possibilities. We demonstrate that these metrics can objectively quantify XAI without user studies and are thus a potentially better way to measure its effectiveness. We provide a use case for how these metrics can be applied to better quantify AI safety and trust within a medical application.

This event is being organised by the King’s Institute for Artificial Intelligence and the UKRI Centre for Doctoral Training in Safe and Trusted Artificial Intelligence.

It is a hybrid event, held at King’s College London and online via Microsoft Teams. Please register with your preferred attendance method via Eventbrite. Students on the STAI CDT do not need to register via Eventbrite.

About the speaker

Avi Rosenfeld (Ph.D. Computer Science, Bar Ilan University, 2007) is an Associate Professor of Computer Science at the Jerusalem College of Technology (JCT). His research generally focuses on data science and artificial intelligence research, and specifically how to reason through machine learning and data mining algorithms to build more accurate and explainable models. His research has created theoretical and practical applications for problems such as robotic coordination, peer to peer full-text search, human-computer interactions, data filtering, scheduling and constraint satisfaction and optimization, communication protocols, scalability issues and medical data analysis. He has published over 80 papers in leading conferences and journals.